Gradient Local Auto Correlation Co-occurrence Machine Learning Model for Endometrial Tuberculosis Identification

Abstract

The effect of endometrial tuberculosis on female fertility is a topic of research for computational aid to assist gynecologists. Early diagnosis using a non-invasive imaging technique, Transvaginal Ultrasound (TVUS) could be of great help. In this paper, a novel effort has been made to effectively classify endometrial TVUS images as normal and abnormal based on the presence of TB. Texture is the most important spatial property in a TVUS image for abnormality identification. The classification approach based on the second-order gray-level co-occurrence matrices (GLCM) extracting statistical texture features are an immense success. As the TVUS images are inherently ill-defined, enhancement of the texture will be beneficial before feature extraction. Since, Gradient Local Auto-Correlation (GLAC) provides enhanced geometric description of gradients and the curvatures of the image surface, GLCM matrix is computed after these features are enhanced by GLAC. Then a hybrid feature selection algorithm is applied on textural features extracted from the GLAC-GLCM matrix, which obtains 66.6% dimensionality reduction and 3.88% increase in classification accuracy. Ultrasound images of female infertility patients have been collected from medical centers using expertise of the medical practitioner in NCR, India. The proposed method on the real dataset obtains a mean accuracy of 85.41% using Support Vector Machine.